4.1. Experiment Originality
The selected point clouds and DEMs presented above, have been produced based on an experiment involving a hybrid UAS and MLS system. The experiment was designed to investigate how the cost-accuracy ratio of UAS photogrammetry could be improved for topographic applications along beach corridors; it was motivated by the interest that a flexible and efficient corridor mapping solution would represent for monitoring long coastal stretches before and after storms.
Presently, the accuracy of UAS-based photogrammetry remains largely dependant on both the acquisition of GCPs and on robust flight geometry (Table 3
]). In terms of field effort, the first issue translates into time-consuming georeferencing of stable ground points [44
] and/or ad-hoc deployment and surveying of optical targets. The second issue requires that images overlap both along and across flight direction. This can be achieved in the case of parallel flight lanes, which at least doubles survey time.
Given these conditions, the experiment design is original, both because of the single-pass flight geometry and because of the reduced number of GCPs, and the MLS data, which is an established technology to survey beach foredunes [20
], could be used as ground truth to validate this innovative survey strategy.
4.2. Hybrid UAS-MLS Performance
For validation purposes, a reference DEM was generated by merging MLS point clouds collected during four passes (P1 to P4 in Table 1
). This reference was then compared to DEMs from individual passes and randomly resampled individual point clouds, attesting that the survey robustness was in line with MLS standards in sandy beach application [20
]. The reference DEM then served to assess the quality of two photogrammetric point clouds generated by integrated sensor orientation
, first with all available GCPs (eighteen) and then with only 4 GCPs and the available KGCPs (231 in total). Both photogrammetically-generated point clouds delivered similar performances, the root mean square of elevation differences with the reference DEM (
) being of 13.8 cm for ISO_18GCPs and of 14.5 cm for ISO_4GCPs+KGCPs.
puts those numbers into perspective with the
values claimed in published applications of photogrammetric topography in sandy coastal settings. In addition to Li et al. [45
], who presented a comparison of the performance for different types of cameras and UAS platforms, the range of values in Table 3
recalls that the quality of photogrammetric products is also sensitive to a wide range of instrumental and environmental parameters. For instance, Gonçalves et al. [40
] and Long et al. [41
] acquired and processed data in a very similar fashion, but still achieved performances which are quite different. Moreover, there is no unique way of comparing point clouds. For instance, Turner et al. [13
] compared the ground elevation, measured with a GNSS receiver mounted on an all-terrain vehicle (ATV), against the average elevation of their high-density photogrammetric point cloud within a horizontal square meter centered on the ATV. Similarly, authors using a greater number of GCPs, surveyed with a pole-mounted GNSS receiver, mostly compared GCPs’ elevation with the elevation interpolated onto the photogrammetrically-derived DEMs [11
]. When both laser (terrestrial or airborne) and photogrammetric data are available, a classic approach is to compare the respectively generated DEMs [27
]. Such comparison implies a densification of the photogrammetric point cloud and, as pointed out by Elsner et al. [27
], may smooth differences compared to the comparison of raw point clouds. Here, the choice was made to quantify the deviation of sparse photogrammetric measurements, from a denoised and a priori
more accurate surface.
Considering these possible technical and methodological differences,
values achieved in this study are very much in line with the expected performance of photogrammetric point clouds. In particular, the performance achieved in this study are better or very close to that of Mancini et al. [39
], Elsner et al. [27
] and Gonçalves et al. [46
] who also compared their photogrammetric point clouds with terrestrial or airborne laser scanner measurements. To be noted also is that
achieved by Laporte-Fauret et al. [44
] with 4 GCPs increases up to 0.5 m to 1.12 m (depending on the camera quality) as they compared their photogrammetric DEMs with 5 across-foredune GNSS profiles.
Likewise, some authors claim performances that are slightly better than those presented herein. For instance, Sturdivant et al. [11
] achieved an umatched
of 3.6 cm, with the lowest above ground flight level (AGL) which also means a lower coverage-to-flight time ratio. Also, increasing the number of GCPs has limits in approving the measurement accuracy. See for instance the relative accuracy in Laporte-Fauret et al. [44
] and in Ruessink et al. [43
] who used a greater number of GCPs. To be noted as well is that the accuracy in Laporte-Fauret et al. [44
] is capped at 5 cm with either 5 or 10 GCPs, similar to results of Tonkin et al. [42
] with either 4 or 101 GCPs. In this respect, mapKITE appears to provide just the required control to match the achievable accuracy, taking into account other instrumental and environmental constraints.
Another important point is that these previous studies rely on more complex flight trajectories, allowing for image overlap in the 2D horizontal plane. Even with differential GNSS processing techniques, it is expected that 2D image overlap should help reduce errors as in the case of Turner et al. [13
]. In the present case, little constraint perpendicularly to the flight direction may have added to the errors and could explain why, even with 18 GCPs,
does not reach the lowest cited values. Indeed, the opposed mean differences on the beach and along the steeper foredune seems to come from a horizontal mismatch or a tilt in the photogrammetric point clouds. Such tilting could be due to less robust image overlaps. Still, with this minimal flight geometry and, in the case ISO_4GCPs+KGCPs, a reduced number of GCPs, observed differences remain within photogrammetric standards.
Regarding the problems that may arise when performing single linear UAS pass, results are good and encouraging. On the beach, the higher variance of the photogrammetric point clouds, compared to the MLS point clouds, contrasts with more similar performances along the foredune slope, and is compensated by extended geodata coverage. The good match between profiles extracted from the DEM produced with the sparse point cloud and those extracted from the reference DEM reveals that photogrammetric point cloud densification was unnecessary to capture the morphological features of the beaches and foredunes.